A closed model is a dish you order at a restaurant: delicious, but you don't know the recipe and you're entirely dependent on the kitchen. An open model is the full recipe handed to you: you can cook it at home, tweak it, share it, and no one can stop you from making it again. In AI, this difference between "ordering the dish" and "having the recipe" decides who truly holds the power.
It's one of the great divides in AI, and it matters more than it looks. On one side, "closed" models like those from OpenAI or Anthropic. On the other, "open" models like many recent Chinese ones. This distinction, often reduced to a technical debate, actually decides questions as broad as national sovereignty and the balance of technological power. Let's break it down.
Closed vs open: the real difference
A closed model (closed source) is only accessible through an online service, usually an API (the interface that lets software talk to the model). You send your request to the company's servers, and it sends back the answer. You never get the model itself in your hands. That's the case with ChatGPT, Claude, and Gemini.
An open model, or more precisely an open-weights model, makes the model's parameters available—those billions of values that make it up. Anyone can download them, run them on their own machines, modify them, and integrate them into their products. The model belongs to you, in the sense that you have a complete, working copy at home.
People often say "open source", but the exact term is usually "open weights". A true open-source software delivers everything needed to rebuild it. Most "open" AI models deliver the final weights, but not always the training data or the full code. That's already huge (you can use and modify the model), but it's not quite the same transparency as a fully free software. The licence accompanying the model, like the famous MIT licence, which is very permissive, spells out what you're allowed to do.
Why it's a question of power
Here's where the technical distinction becomes geopolitical. A closed model can be cut off from you. The company can change its prices, alter its terms, or be forced by a government to suspend access, as Anthropic experienced when the US state temporarily blocked Fable 5. You depend on a third party that keeps the upper hand.
An open-weights model, once downloaded, can't be switched off by anyone. No corporate or government decision can take away a copy you already own. That's exactly the strategic argument put forward by Chinese labs like Zhipu or Meituan, whose rise we covered with LongCat-2.0. Facing US restrictions, openness becomes a weapon: a model that can't be cut off is a model free from any political pressure.
The two sides of the coin
Openness has obvious virtues. It democratises access to AI, lets researchers and small companies experiment without paying for an API, guarantees confidentiality (your data never leaves your machines), and offers precious independence. For a regulated sector, a government body, or a company concerned about sovereignty, being able to host your own AI is a major asset.
But openness also has a downside. A freely downloadable model can be used for malicious purposes with no safeguards, since anyone can strip out the safety measures placed on it. That's the whole dilemma: the very property that makes an open model impossible to censor also makes it impossible to control. Labs that keep their models closed often cite this security reason, on top of the obvious commercial one.
What to take away
Open or closed isn't just a technical choice; it's a societal one. Closed models offer more control and security, but create dependency. Open models offer freedom and sovereignty, but escape any centralised control. Both approaches coexist and respond to each other, and the tension between them will shape AI for years to come.
For you, the right question isn't "which is better?", but "which fits my need?". For simple consumer use, a well-integrated closed model is more than enough. For a company determined not to depend on anyone, for a government body attached to its sovereignty, or for anyone who wants to keep their data strictly at home, open source becomes a valuable—even essential—option. Understanding this distinction is understanding one of the most structuring choices of the entire AI era.